-
Notifications
You must be signed in to change notification settings - Fork 6
/
eval_pck.py
242 lines (221 loc) · 12.8 KB
/
eval_pck.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
import time
import sys, os
import pdb
import torch
import torch.nn as nn
from torch.autograd import Variable
sys.path.insert(0,'third_party')
from ext_utils.badja_data import BADJAData
from ext_utils.joint_catalog import SMALJointInfo
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn.functional as F
import cv2
import pdb
import soft_renderer as sr
import argparse
import trimesh
from nnutils.geom_utils import obj_to_cam, pinhole_cam, orthographic_cam, render_flow_soft_3
parser = argparse.ArgumentParser(description='BADJA')
parser.add_argument('--testdir', default='',
help='path to test dir')
parser.add_argument('--seqname', default='camel',
help='sequence to test')
parser.add_argument('--type', default='mesh',
help='load mesh data or flow or zero')
parser.add_argument('--cam_type', default='perspective',
help='camera model, orthographic or perspective')
parser.add_argument('--vis', default='no',
help='whether to draw visualization')
parser.add_argument('--cse_mesh_name', default='smpl_27554',
help='whether to draw visualization')
args = parser.parse_args()
renderer_softflf = sr.SoftRenderer(image_size=256,dist_func='hard' ,aggr_func_alpha='hard',
camera_mode='look_at',perspective=False, aggr_func_rgb='hard',
light_mode='vertex', light_intensity_ambient=1.,light_intensity_directionals=0.)
def preprocess_image(img,mask,imgsize):
if len(img.shape) == 2:
img = np.repeat(np.expand_dims(img, 2), 3, axis=2)
if mask.shape[0]!=img.shape[0] or mask.shape[1]!=img.shape[1]:
mask = cv2.resize(mask, img.shape[:2][::-1],interpolation=cv2.INTER_NEAREST)[:,:,None]
# crop box
indices = np.where(mask>0); xid = indices[1]; yid = indices[0]
center = ( (xid.max()+xid.min())//2, (yid.max()+yid.min())//2)
length = ( (xid.max()-xid.min())//2, (yid.max()-yid.min())//2)
length = (int(1.2*length[0]), int(1.2*length[1]))
alp = (2*length[0]/float(imgsize), 2*length[1]/float(imgsize))
refpp = np.asarray(center)/(imgsize/2.) - 1
return alp, refpp,center,length[0]
def draw_joints_on_image(rgb_img, joints, visibility, region_colors, marker_types,pred=None,correct=None):
joints = joints[:, ::-1] # OpenCV works in (x, y) rather than (i, j)
disp_img = rgb_img.copy()
i=0
for joint_coord, visible, color, marker_type in zip(joints, visibility, region_colors, marker_types):
if visible:
joint_coord = joint_coord.astype(int)
cv2.circle(disp_img, tuple(joint_coord), radius=3, color=[255,0,0], thickness = 10)
if pred is not None:
if correct[i]:
color=[0,255,0]
else:
color=[0,0,255]
error = np.linalg.norm(joint_coord - pred[i,::-1],2,-1)
cv2.circle(disp_img, tuple(joint_coord), radius=int(error), color=color, thickness = 3)
cv2.line(disp_img, tuple(joint_coord), tuple(pred[i,::-1]),color , thickness = 3)
i+=1
return disp_img
def main():
smal_joint_info = SMALJointInfo()
badja_data = BADJAData(args.seqname)
data_loader = badja_data.get_loader()
print(args.testdir)
# store all the data
all_anno = []
all_mesh = []
all_cam = []
all_fr = []
all_fl = []
for anno in data_loader:
all_anno.append(anno)
rgb_img, sil_img, joints, visible, name = anno
seqname = name.split('/')[-2]
fr = int(name.split('/')[-1].split('.')[-2])
all_fr.append(fr)
print('%s/%d'%(seqname, fr))
# load mesh data or flow
if args.type=='mesh':
if os.path.exists('%s/%s-pred%d.ply'%(args.testdir, seqname, fr)):
mesh = trimesh.load('%s/%s-pred%d.ply'%(args.testdir, seqname, fr),process=False)
cam = np.loadtxt('%s/%s-cam%d.txt'%(args.testdir,seqname, fr))
elif os.path.exists('%s/r%s-pred%d.ply'%(args.testdir, seqname, fr)):
mesh = trimesh.load('%s/r%s-pred%d.ply'%(args.testdir, seqname, fr),process=False)
cam = np.loadtxt('%s/r%s-cam%d.txt'%(args.testdir,seqname, fr))
else:
print('data not found')
exit()
all_mesh.append(mesh)
all_cam.append(cam)
if args.type=='flow':
import sys
sys.path.insert(0,'data_gen')
from models.VCNplus import VCN
model = VCN([1, 256, 256], md=[int(4*(256/256)),4,4,4,4], fac=1)
model = nn.DataParallel(model, device_ids=[0])
model.cuda()
pretrained_dict = torch.load('/data/gengshay/lasr_vcn/flow-rob-4th.pth',map_location='cpu')
mean_L=pretrained_dict['mean_L']
mean_R=pretrained_dict['mean_R']
model.load_state_dict(pretrained_dict['state_dict'],strict=False)
if args.type=='cse':
import sys
sys.path.insert(0,'/data/gengshay/code/detectron2_previous/projects/DensePose')
from cselib import create_cse, run_cse
if args.cse_mesh_name=='smpl_27554':
model, embedder, mesh_vertex_embeddings = create_cse(isanimal=False)
else:
model, embedder, mesh_vertex_embeddings = create_cse(isanimal=True)
import pickle
with open('/data/gengshay/code/detectron2_previous/projects/DensePose/geodists_%s.pkl'%(args.cse_mesh_name), 'rb') as f: geodists=pickle.load(f)
geodists = torch.Tensor(geodists).cuda()
geodists[0,:] = np.inf
geodists[:,0] = np.inf
# store all the results
pck_all = []
for i in range(len(all_anno)):
for j in range(len(all_anno)):
if i!=j:
# evaluate every two-frame
refimg, refsil, refkp, refvis, refname = all_anno[i]
tarimg, tarsil, tarkp, tarvis, tarname = all_anno[j]
refseqname = refname.split('/')[-2]
tarseqname = tarname.split('/')[-2]
# control inter or inner; comment to compute inter
#if refseqname!= tarseqname: continue
print('%s vs %s'%(refname, tarname))
if args.type=='mesh':
refmesh, tarmesh = all_mesh[i], all_mesh[j]
refcam, tarcam = all_cam[i], all_cam[j]
img_size = max(refimg.shape)
reffl, refpp, center, length = preprocess_image(refimg,refsil,img_size)
reffl2, refpp2, center2, length2 = preprocess_image(tarimg,tarsil,img_size)
reffl = np.stack([reffl,reffl2])
refpp = np.stack([refpp,refpp2])
renderer_softflf.rasterizer.image_size = img_size
# render flow between mesh 1 and 2
refface = torch.Tensor(refmesh.faces[None]).cuda()
verts = torch.Tensor(np.concatenate([refmesh.vertices[None], tarmesh.vertices[None]],0)).cuda()
Rmat = torch.Tensor(np.concatenate([refcam[None,:3,:3], tarcam[None,:3,:3]], 0)).cuda()
Tmat = torch.Tensor(np.concatenate([refcam[None,:3,3], tarcam[None,:3,3]], 0)).cuda()
ppoint = torch.Tensor(np.concatenate([refcam[None,3,2:], tarcam[None,3,2:]], 0)).cuda()
scale = torch.Tensor(np.concatenate([refcam[None,3,:2], tarcam[None,3,:2]], 0)).cuda()
scale = scale/img_size*2
ppoint = ppoint/img_size * 2 -1
verts_fl = obj_to_cam(verts, Rmat, Tmat[:,None],nmesh=1,n_hypo=1,skin=None)
verts_fl = torch.cat([verts_fl,torch.ones_like(verts_fl[:, :, 0:1])], dim=-1)
verts_pos = verts_fl.clone()
if args.cam_type=='perspective':
verts_fl = pinhole_cam(verts_fl, ppoint, scale)
flow_fw, bgmask_fw, fgmask_flowf = render_flow_soft_3(renderer_softflf, verts_fl[:1], verts_fl[1:], refface)
elif args.cam_type=='orthographic':
verts_fl = orthographic_cam(verts_fl, ppoint, scale)
verts_fl[:,:,-2] += (-verts_fl[:,:,-2].min()+10)
renderer_softflf.rasterizer.far=verts_fl[:,:,2].max() + 10
renderer_softflf.rasterizer.near=verts_fl[:,:,2].min() - 10
flow_fw, bgmask_fw, fgmask_flowf = render_flow_soft_3(renderer_softflf, verts_fl[:1], verts_fl[1:], refface)
else:exit()
flow_fw[bgmask_fw]=0.
flow_fw[:,:,:,0] *= flow_fw.shape[2] / 2
flow_fw[:,:,:,1] *= flow_fw.shape[1] / 2
flow_fw = torch.cat([flow_fw, torch.zeros_like(flow_fw)[:,:,:,:1]],-1)[:,:refimg.shape[0],:refimg.shape[1]]
elif args.type=='flow':
flow_fw = process_flow(model, refimg, tarimg, mean_L, mean_R)[0]
flow_fw = torch.Tensor(flow_fw[None]).cuda()
elif args.type=='zero':
flow_fw = torch.zeros(refimg.shape).cuda()[None]
elif args.type=='cse':
imh, imw = refimg.shape[:2]
csmfw1, csmbw1, image_bgr1, bbox1,bbl1 = run_cse(model, embedder, mesh_vertex_embeddings, refimg[:,:,::-1], refsil[:,:,0], mesh_name=args.cse_mesh_name)
csmfw2, csmbw2, image_bgr2, bbox2,bbl2 = run_cse(model, embedder, mesh_vertex_embeddings, tarimg[:,:,::-1], tarsil[:,:,0], mesh_name=args.cse_mesh_name)
csmfw1 = csmfw1.view(-1,1).repeat(1,112*112).view(-1,1) # dismat is f1xf2
csmfw2 = csmfw2.view(1,-1).repeat(112*112,1).view(-1,1)
idxmatch = geodists[csmfw1, csmfw2].view(112*112, 112*112).argmin(1)
tar_coord = torch.cat([idxmatch[:,None]%112, idxmatch[:,None]//112],-1).float() # cx,cy
tar_coord[:,0]=tar_coord[:,0]*bbl2[0]/112 + bbox2[0]
tar_coord[:,1]=tar_coord[:,1]*bbl2[1]/112 + bbox2[1]
tar_coord = tar_coord.view(112, 112, 2)
ref_coord = torch.Tensor(np.meshgrid(range(112), range(112))).cuda().permute(1,2,0).view(-1,2)
ref_coord[:,0]=ref_coord[:,0]*bbl1[0]/112 + bbox1[0]
ref_coord[:,1]=ref_coord[:,1]*bbl1[1]/112 + bbox1[1]
ref_coord = ref_coord.view(112, 112, 2)
flow_fw = tar_coord - ref_coord
flow_fw = F.interpolate(flow_fw.permute(2,0,1)[None], [bbl1[1],bbl1[0]], mode='bilinear')[0].permute(1,2,0)
tmp = torch.zeros(refimg.shape[0], refimg.shape[1],3).cuda()
tmp[bbox1[1]:bbox1[3],bbox1[0]:bbox1[2],:2] = flow_fw
flow_fw = tmp
flow_fw[torch.Tensor(refsil).cuda().repeat(1,1,3)<=0] = 0
flow_fw = flow_fw[None]
refkpx = torch.Tensor(refkp.astype(float)).cuda()
x0,y0=np.meshgrid(range(refimg.shape[1]),range(refimg.shape[0]))
x0 = torch.Tensor(x0).cuda()
y0 = torch.Tensor(y0).cuda()
idx = ( (flow_fw[:,:,:,:2].norm(2,-1)<1e-6).float().view(1,-1)*1e6+ (torch.pow(refkpx[:,0:1]-y0.view(1,-1),2) + torch.pow(refkpx[:,1:2]-x0.view(1,-1),2))/1000 ).argmin(-1)
samp_flow = flow_fw.view(-1,3)[idx][:,:2]
tarkp_pred = refkpx.clone()
tarkp_pred[:,0] = tarkp_pred[:,0] +(samp_flow[:,1])
tarkp_pred[:,1] = tarkp_pred[:,1] +(samp_flow[:,0])
tarkp_pred = np.asarray(tarkp_pred.cpu())
diff = np.linalg.norm(tarkp_pred - tarkp, 2,-1)
sqarea = np.sqrt((refsil[:,:,0]>0).sum())
correct = diff < sqarea * 0.2
correct = correct[np.logical_and(tarvis, refvis)]
if args.vis=='yes':
rgb_vis = draw_joints_on_image(refimg, refkp, refvis, smal_joint_info.joint_colors, smal_joint_info.annotated_markers)
tarimg = draw_joints_on_image(tarimg, tarkp, refvis, smal_joint_info.joint_colors, smal_joint_info.annotated_markers, pred=tarkp_pred,correct=diff < sqarea * 0.2)
cv2.imwrite('%s/%s-%05d-%s-%05d-ref.png'%(args.testdir,refseqname, all_fr[i],tarseqname,all_fr[j]),rgb_vis[:,:,::-1])
cv2.imwrite('%s/%s-%05d-%s-%05d-tar.png'%(args.testdir, refseqname, all_fr[i],tarseqname,all_fr[j]),tarimg[:,:,::-1])
write_pfm( '%s/%s-%05d-%s-%05d.pfm'%(args.testdir, refseqname, all_fr[i], tarseqname, all_fr[j]),np.asarray(flow_fw[0].detach().cpu()))
pck_all.append(correct)
print('PCK %.02f'%(100*np.concatenate(pck_all).astype(float).mean()))
if __name__ == '__main__':
main()